Commit | Line | Data |
---|---|---|
fb7b71a3 JA |
1 | /* |
2 | * blktrace support code for fio | |
3 | */ | |
4 | #include <stdio.h> | |
5 | #include <stdlib.h> | |
5e6c2067 JA |
6 | #include <sys/stat.h> |
7 | #include <dirent.h> | |
8c1fdf04 | 8 | |
01743ee1 | 9 | #include "flist.h" |
fb7b71a3 JA |
10 | #include "fio.h" |
11 | #include "blktrace_api.h" | |
12 | ||
2da7df1d | 13 | #define TRACE_FIFO_SIZE 8192 |
e2887563 JA |
14 | |
15 | /* | |
16 | * fifo refill frontend, to avoid reading data in trace sized bites | |
17 | */ | |
18 | static int refill_fifo(struct thread_data *td, struct fifo *fifo, int fd) | |
19 | { | |
20 | char buf[TRACE_FIFO_SIZE]; | |
f12b323f | 21 | unsigned int total; |
e2887563 JA |
22 | int ret; |
23 | ||
f12b323f JA |
24 | total = sizeof(buf); |
25 | if (total > fifo_room(fifo)) | |
26 | total = fifo_room(fifo); | |
e2887563 | 27 | |
f12b323f JA |
28 | ret = read(fd, buf, total); |
29 | if (ret < 0) { | |
30 | td_verror(td, errno, "read blktrace file"); | |
31 | return -1; | |
e2887563 JA |
32 | } |
33 | ||
f12b323f JA |
34 | if (ret > 0) |
35 | ret = fifo_put(fifo, buf, ret); | |
36 | ||
bd6f78b2 | 37 | dprint(FD_BLKTRACE, "refill: filled %d bytes\n", ret); |
f12b323f | 38 | return ret; |
e2887563 JA |
39 | } |
40 | ||
41 | /* | |
42 | * Retrieve 'len' bytes from the fifo, refilling if necessary. | |
43 | */ | |
44 | static int trace_fifo_get(struct thread_data *td, struct fifo *fifo, int fd, | |
45 | void *buf, unsigned int len) | |
46 | { | |
f12b323f JA |
47 | if (fifo_len(fifo) < len) { |
48 | int ret = refill_fifo(td, fifo, fd); | |
e2887563 | 49 | |
f12b323f JA |
50 | if (ret < 0) |
51 | return ret; | |
52 | } | |
e2887563 JA |
53 | |
54 | return fifo_get(fifo, buf, len); | |
55 | } | |
56 | ||
8c1fdf04 JA |
57 | /* |
58 | * Just discard the pdu by seeking past it. | |
59 | */ | |
f12b323f JA |
60 | static int discard_pdu(struct thread_data *td, struct fifo *fifo, int fd, |
61 | struct blk_io_trace *t) | |
fb7b71a3 JA |
62 | { |
63 | if (t->pdu_len == 0) | |
64 | return 0; | |
65 | ||
bd6f78b2 | 66 | dprint(FD_BLKTRACE, "discard pdu len %u\n", t->pdu_len); |
f12b323f | 67 | return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len); |
fb7b71a3 JA |
68 | } |
69 | ||
8c1fdf04 JA |
70 | /* |
71 | * Check if this is a blktrace binary data file. We read a single trace | |
72 | * into memory and check for the magic signature. | |
73 | */ | |
fb7b71a3 JA |
74 | int is_blktrace(const char *filename) |
75 | { | |
76 | struct blk_io_trace t; | |
77 | int fd, ret; | |
78 | ||
79 | fd = open(filename, O_RDONLY); | |
4dced407 | 80 | if (fd < 0) |
fb7b71a3 | 81 | return 0; |
fb7b71a3 JA |
82 | |
83 | ret = read(fd, &t, sizeof(t)); | |
84 | close(fd); | |
85 | ||
86 | if (ret < 0) { | |
87 | perror("read blktrace"); | |
88 | return 0; | |
89 | } else if (ret != sizeof(t)) { | |
90 | log_err("fio: short read on blktrace file\n"); | |
91 | return 0; | |
92 | } | |
93 | ||
94 | if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) | |
95 | return 1; | |
96 | ||
97 | return 0; | |
98 | } | |
99 | ||
d1c46c04 DN |
100 | static int lookup_device(struct thread_data *td, char *path, unsigned int maj, |
101 | unsigned int min) | |
5e6c2067 JA |
102 | { |
103 | struct dirent *dir; | |
104 | struct stat st; | |
105 | int found = 0; | |
106 | DIR *D; | |
107 | ||
108 | D = opendir(path); | |
109 | if (!D) | |
110 | return 0; | |
111 | ||
112 | while ((dir = readdir(D)) != NULL) { | |
113 | char full_path[256]; | |
114 | ||
115 | if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, "..")) | |
116 | continue; | |
117 | ||
b9fd788f | 118 | sprintf(full_path, "%s%s%s", path, FIO_OS_PATH_SEPARATOR, dir->d_name); |
5e6c2067 JA |
119 | if (lstat(full_path, &st) == -1) { |
120 | perror("lstat"); | |
121 | break; | |
122 | } | |
123 | ||
124 | if (S_ISDIR(st.st_mode)) { | |
d1c46c04 | 125 | found = lookup_device(td, full_path, maj, min); |
5e6c2067 JA |
126 | if (found) { |
127 | strcpy(path, full_path); | |
128 | break; | |
129 | } | |
130 | } | |
131 | ||
132 | if (!S_ISBLK(st.st_mode)) | |
133 | continue; | |
134 | ||
d1c46c04 DN |
135 | /* |
136 | * If replay_redirect is set then always return this device | |
137 | * upon lookup which overrides the device lookup based on | |
138 | * major minor in the actual blktrace | |
139 | */ | |
140 | if (td->o.replay_redirect) { | |
141 | dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden" | |
142 | " with: %s", maj, min, | |
143 | td->o.replay_redirect); | |
144 | strcpy(path, td->o.replay_redirect); | |
145 | found = 1; | |
146 | break; | |
147 | } | |
148 | ||
5e6c2067 | 149 | if (maj == major(st.st_rdev) && min == minor(st.st_rdev)) { |
5ec10eaa | 150 | dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min); |
5e6c2067 JA |
151 | strcpy(path, full_path); |
152 | found = 1; | |
153 | break; | |
154 | } | |
155 | } | |
156 | ||
157 | closedir(D); | |
158 | return found; | |
159 | } | |
160 | ||
c69aa91f JA |
161 | #define FMINORBITS 20 |
162 | #define FMINORMASK ((1U << FMINORBITS) - 1) | |
163 | #define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS)) | |
164 | #define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK)) | |
eeb9c2aa | 165 | |
691c8fb0 JA |
166 | static void trace_add_open_event(struct thread_data *td, int fileno) |
167 | { | |
168 | struct io_piece *ipo; | |
169 | ||
170 | ipo = calloc(1, sizeof(*ipo)); | |
0d29de83 | 171 | init_ipo(ipo); |
691c8fb0 JA |
172 | |
173 | ipo->ddir = DDIR_INVAL; | |
174 | ipo->fileno = fileno; | |
175 | ipo->file_action = FIO_LOG_OPEN_FILE; | |
01743ee1 | 176 | flist_add_tail(&ipo->list, &td->io_log_list); |
691c8fb0 JA |
177 | } |
178 | ||
5e6c2067 JA |
179 | static void trace_add_file(struct thread_data *td, __u32 device) |
180 | { | |
181 | static unsigned int last_maj, last_min; | |
c69aa91f JA |
182 | unsigned int maj = FMAJOR(device); |
183 | unsigned int min = FMINOR(device); | |
5e6c2067 JA |
184 | struct fio_file *f; |
185 | char dev[256]; | |
186 | unsigned int i; | |
187 | ||
188 | if (last_maj == maj && last_min == min) | |
189 | return; | |
190 | ||
191 | last_maj = maj; | |
192 | last_min = min; | |
193 | ||
194 | /* | |
195 | * check for this file in our list | |
196 | */ | |
197 | for_each_file(td, f, i) | |
198 | if (f->major == maj && f->minor == min) | |
199 | return; | |
200 | ||
201 | strcpy(dev, "/dev"); | |
d1c46c04 | 202 | if (lookup_device(td, dev, maj, min)) { |
691c8fb0 JA |
203 | int fileno; |
204 | ||
bd6f78b2 | 205 | dprint(FD_BLKTRACE, "add devices %s\n", dev); |
49ffb4a2 | 206 | fileno = add_file_exclusive(td, dev); |
691c8fb0 | 207 | trace_add_open_event(td, fileno); |
bd6f78b2 | 208 | } |
5e6c2067 JA |
209 | } |
210 | ||
8c1fdf04 JA |
211 | /* |
212 | * Store blk_io_trace data in an ipo for later retrieval. | |
213 | */ | |
fdefd987 | 214 | static void store_ipo(struct thread_data *td, unsigned long long offset, |
8c1fdf04 | 215 | unsigned int bytes, int rw, unsigned long long ttime) |
fdefd987 JA |
216 | { |
217 | struct io_piece *ipo = malloc(sizeof(*ipo)); | |
218 | ||
0d29de83 JA |
219 | init_ipo(ipo); |
220 | ||
a2eea81b JA |
221 | /* |
222 | * the 512 is wrong here, it should be the hardware sector size... | |
223 | */ | |
224 | ipo->offset = offset * 512; | |
fdefd987 | 225 | ipo->len = bytes; |
8c1fdf04 | 226 | ipo->delay = ttime / 1000; |
fdefd987 JA |
227 | if (rw) |
228 | ipo->ddir = DDIR_WRITE; | |
229 | else | |
230 | ipo->ddir = DDIR_READ; | |
231 | ||
bd6f78b2 JA |
232 | dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n", |
233 | ipo->ddir, ipo->offset, | |
234 | ipo->len, ipo->delay); | |
691c8fb0 | 235 | queue_io_piece(td, ipo); |
fdefd987 JA |
236 | } |
237 | ||
0b9d69ec | 238 | static void handle_trace_notify(struct blk_io_trace *t) |
cd991b9e | 239 | { |
691c8fb0 JA |
240 | switch (t->action) { |
241 | case BLK_TN_PROCESS: | |
242 | printf("got process notify: %x, %d\n", t->action, t->pid); | |
243 | break; | |
244 | case BLK_TN_TIMESTAMP: | |
245 | printf("got timestamp notify: %x, %d\n", t->action, t->pid); | |
246 | break; | |
ff58fced JA |
247 | case BLK_TN_MESSAGE: |
248 | break; | |
691c8fb0 JA |
249 | default: |
250 | dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action); | |
251 | break; | |
252 | } | |
253 | } | |
5b3023b8 | 254 | |
ff58fced JA |
255 | static void handle_trace_discard(struct thread_data *td, struct blk_io_trace *t, |
256 | unsigned long long ttime, unsigned long *ios) | |
257 | { | |
258 | struct io_piece *ipo = malloc(sizeof(*ipo)); | |
259 | ||
0d29de83 | 260 | init_ipo(ipo); |
ff58fced JA |
261 | trace_add_file(td, t->device); |
262 | ||
263 | ios[DDIR_WRITE]++; | |
264 | td->o.size += t->bytes; | |
265 | ||
266 | memset(ipo, 0, sizeof(*ipo)); | |
267 | INIT_FLIST_HEAD(&ipo->list); | |
268 | ||
269 | /* | |
270 | * the 512 is wrong here, it should be the hardware sector size... | |
271 | */ | |
272 | ipo->offset = t->sector * 512; | |
273 | ipo->len = t->bytes; | |
274 | ipo->delay = ttime / 1000; | |
275 | ipo->ddir = DDIR_TRIM; | |
276 | ||
277 | dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n", | |
278 | ipo->offset, ipo->len, | |
279 | ipo->delay); | |
280 | queue_io_piece(td, ipo); | |
281 | } | |
282 | ||
691c8fb0 JA |
283 | static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t, |
284 | unsigned long long ttime, unsigned long *ios, | |
285 | unsigned int *bs) | |
286 | { | |
287 | int rw; | |
5b3023b8 JA |
288 | |
289 | trace_add_file(td, t->device); | |
290 | ||
291 | rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0; | |
292 | ||
293 | if (t->bytes > bs[rw]) | |
294 | bs[rw] = t->bytes; | |
295 | ||
296 | ios[rw]++; | |
297 | td->o.size += t->bytes; | |
298 | store_ipo(td, t->sector, t->bytes, rw, ttime); | |
cd991b9e JA |
299 | } |
300 | ||
691c8fb0 JA |
301 | /* |
302 | * We only care for queue traces, most of the others are side effects | |
303 | * due to internal workings of the block layer. | |
304 | */ | |
305 | static void handle_trace(struct thread_data *td, struct blk_io_trace *t, | |
306 | unsigned long long ttime, unsigned long *ios, | |
307 | unsigned int *bs) | |
308 | { | |
309 | if ((t->action & 0xffff) != __BLK_TA_QUEUE) | |
310 | return; | |
311 | if (t->action & BLK_TC_ACT(BLK_TC_PC)) | |
312 | return; | |
313 | ||
314 | if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY)) | |
aec2de20 | 315 | handle_trace_notify(t); |
ff58fced JA |
316 | else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD)) |
317 | handle_trace_discard(td, t, ttime, ios); | |
691c8fb0 JA |
318 | else |
319 | handle_trace_fs(td, t, ttime, ios, bs); | |
320 | } | |
321 | ||
8c1fdf04 JA |
322 | /* |
323 | * Load a blktrace file by reading all the blk_io_trace entries, and storing | |
324 | * them as io_pieces like the fio text version would do. | |
325 | */ | |
fb7b71a3 JA |
326 | int load_blktrace(struct thread_data *td, const char *filename) |
327 | { | |
a61eddec | 328 | unsigned long long ttime, delay; |
fb7b71a3 | 329 | struct blk_io_trace t; |
4241ea8f | 330 | unsigned long ios[2], skipped_writes; |
a61eddec | 331 | unsigned int cpu; |
d84f8d49 | 332 | unsigned int rw_bs[2]; |
e2887563 | 333 | struct fifo *fifo; |
fb7b71a3 JA |
334 | int fd; |
335 | ||
336 | fd = open(filename, O_RDONLY); | |
337 | if (fd < 0) { | |
338 | td_verror(td, errno, "open blktrace file"); | |
339 | return 1; | |
340 | } | |
341 | ||
e2887563 JA |
342 | fifo = fifo_alloc(TRACE_FIFO_SIZE); |
343 | ||
6df8adaa JA |
344 | td->o.size = 0; |
345 | ||
a61eddec | 346 | cpu = 0; |
d84f8d49 JA |
347 | ttime = 0; |
348 | ios[0] = ios[1] = 0; | |
349 | rw_bs[0] = rw_bs[1] = 0; | |
4241ea8f | 350 | skipped_writes = 0; |
fb7b71a3 | 351 | do { |
e2887563 | 352 | int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t)); |
fb7b71a3 | 353 | |
e2887563 | 354 | if (ret < 0) |
8c1fdf04 | 355 | goto err; |
e2887563 JA |
356 | else if (!ret) |
357 | break; | |
358 | else if (ret < (int) sizeof(t)) { | |
359 | log_err("fio: short fifo get\n"); | |
fb7b71a3 | 360 | break; |
fb7b71a3 JA |
361 | } |
362 | ||
363 | if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) { | |
5ec10eaa JA |
364 | log_err("fio: bad magic in blktrace data: %x\n", |
365 | t.magic); | |
8c1fdf04 | 366 | goto err; |
fb7b71a3 JA |
367 | } |
368 | if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) { | |
5ec10eaa JA |
369 | log_err("fio: bad blktrace version %d\n", |
370 | t.magic & 0xff); | |
8c1fdf04 | 371 | goto err; |
fb7b71a3 | 372 | } |
f12b323f JA |
373 | ret = discard_pdu(td, fifo, fd, &t); |
374 | if (ret < 0) { | |
fb7b71a3 | 375 | td_verror(td, ret, "blktrace lseek"); |
8c1fdf04 | 376 | goto err; |
f12b323f JA |
377 | } else if (t.pdu_len != ret) { |
378 | log_err("fio: discarded %d of %d\n", ret, t.pdu_len); | |
379 | goto err; | |
fb7b71a3 | 380 | } |
691c8fb0 JA |
381 | if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) { |
382 | if (!ttime) { | |
383 | ttime = t.time; | |
384 | cpu = t.cpu; | |
385 | } | |
386 | ||
387 | delay = 0; | |
388 | if (cpu == t.cpu) | |
389 | delay = t.time - ttime; | |
390 | if ((t.action & BLK_TC_ACT(BLK_TC_WRITE)) && read_only) | |
391 | skipped_writes++; | |
64bbb865 DN |
392 | else { |
393 | /* | |
394 | * set delay to zero if no_stall enabled for | |
395 | * fast replay | |
396 | */ | |
397 | if (td->o.no_stall) | |
398 | delay = 0; | |
399 | ||
691c8fb0 | 400 | handle_trace(td, &t, delay, ios, rw_bs); |
64bbb865 | 401 | } |
691c8fb0 | 402 | |
8c1fdf04 | 403 | ttime = t.time; |
a61eddec | 404 | cpu = t.cpu; |
a6edd638 JA |
405 | } else { |
406 | delay = 0; | |
4241ea8f | 407 | handle_trace(td, &t, delay, ios, rw_bs); |
a6edd638 | 408 | } |
fb7b71a3 JA |
409 | } while (1); |
410 | ||
38470f85 | 411 | fifo_free(fifo); |
fb7b71a3 | 412 | close(fd); |
8c1fdf04 | 413 | |
4241ea8f | 414 | if (skipped_writes) |
5ec10eaa JA |
415 | log_err("fio: %s skips replay of %lu writes due to read-only\n", |
416 | td->o.name, skipped_writes); | |
4241ea8f | 417 | |
8c1fdf04 JA |
418 | if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) { |
419 | log_err("fio: found no ios in blktrace data\n"); | |
420 | return 1; | |
d84f8d49 | 421 | } else if (ios[DDIR_READ] && !ios[DDIR_READ]) { |
8c1fdf04 | 422 | td->o.td_ddir = TD_DDIR_READ; |
d84f8d49 JA |
423 | td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ]; |
424 | } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) { | |
8c1fdf04 | 425 | td->o.td_ddir = TD_DDIR_WRITE; |
d84f8d49 JA |
426 | td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE]; |
427 | } else { | |
8c1fdf04 | 428 | td->o.td_ddir = TD_DDIR_RW; |
d84f8d49 JA |
429 | td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ]; |
430 | td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE]; | |
431 | } | |
8c1fdf04 JA |
432 | |
433 | /* | |
434 | * We need to do direct/raw ios to the device, to avoid getting | |
435 | * read-ahead in our way. | |
436 | */ | |
437 | td->o.odirect = 1; | |
438 | ||
fb7b71a3 | 439 | return 0; |
8c1fdf04 JA |
440 | err: |
441 | close(fd); | |
38470f85 | 442 | fifo_free(fifo); |
8c1fdf04 | 443 | return 1; |
fb7b71a3 | 444 | } |